Submission v2: arthurmichel00 — 128.79 official train / 130.25 holdout (supersedes #51)

[arthurmichel00]

MIB Doc Challenge Submission (v2 — supersedes #51)

Per the organizers' guidance that updated submissions are filed as a new PR: this replaces #51 (same author). Delta vs #51: one fully-gated fix — a truncation-ambiguity guard in the trusted-Finding fuzzy matcher (a damage-truncated "NEEDS_REVIEW" could previously mis-bind to DENIED under the image's Tesseract). Shipped only after the complete battery: test suite, red-team 4/4, trap corpus 6/6, and a 23,066-line matcher sweep with both changed lines individually verified. It changes zero verdicts on the 5,000 validation cases; the measured gain is on train.

Links

• Solution repository (public, contains a Dockerfile): https://github.com/arthurmichel00/mib-doc-solution
• Submission form: submitted 2026-08-02 with this PR's link (earlier entry for Submission: arthurmichel00 — measured, fail-closed document pipeline (128.69 official train / 130.25 holdout) #51 filed 2026-08-01).

Summary

Official Docker-contract train score: 128.79 / 150 · fixed 200-case holdout: 130.25 (holdout ≥ train at all 8 measured milestones) · exactly 1 catastrophic false approval (the documented designed trap) · 0 fallback rows · 5.64 s/PDF vs 6.0 budget · 0.36 GiB image. predictions.jsonl generated by the frozen image over all 5,000 validation PDFs under the offline contract; passes validate_submission.py (5,000 valid, 0 missing). Approved-side consistency enforced structurally at emission and separately re-audited: 0 violations. No answer-key ingestion — the planted key is adversarial and never read as evidence. Built by AI agents in gated autonomous loops (30+ levers A/B-measured, 10 shipped; receipts in the repo's LEVERS.md and APPENDIX.md).

Checklist

• I filled out the submission form linked above (submitted 2026-08-02 with this PR's link; Submission: arthurmichel00 — measured, fail-closed document pipeline (128.69 official train / 130.25 holdout) #51's earlier form entry also on file)
• This PR only adds submissions/arthurmichel00/predictions.jsonl, MEMO.md, and SUBMISSION.md
• predictions.jsonl passes scripts/validate_submission.py against data/validation_manifest.csv
• My solution repository is public and includes a Dockerfile
• My Docker image runs offline (--network none) and accepts <input_pdf_dir> <output_predictions_path>
• My submitted runtime uses no LLMs, VLMs, cloud OCR, or network services
• Model artifacts fit the size limits in DOCKER_SUBMISSION.md
• No hardcoded validation answers and no manual per-case edits
• My memo describes my approach, failure modes, and what I would improve with another week

🤖 Generated with Claude Code

[arthurmichel00]

Superseded by #68 (same author) — v3 adds a constrained-candidate CTC fill channel and three decode-repair mechanisms, measured +0.26 official train / +0.25 holdout, with every adjudication flip gold-verified and the catastrophic false-approval count unchanged.